Abstract
Algorithmic decision support is rapidly becoming a staple of personalized medicine, particularly for high-stakes recommendations such as cancer subtyping in which access to patient-specific information can drastically alter the course of treatment, and thus, patient outcome. To enhance the utility of decision support systems, it is vital to provide not just recommendations, but also contextual information that aids human understanding and improves collaboration between humans and machines. This paper introduces an automated, explainable system designed to assist pathologists with cancer subtyping from digital cytopathology images, incorporating interactive content inspection for quick assessment of cell composition. Adhering to human-centered design principles, our approach draws from a formative study to mirror pathologists' examination processes in the algorithm's design. Our deep learning-based method first analyzes every cell in the cytopathology slide to extract and cluster their appearance, and then uses an interpretable rule set to determine cancer subtypes based on overall cell composition clusters. This method enables verification of cell composition through interactive inspection and offers transparent decision rules, enhancing human-machine collaboration without sacrificing performance compared to traditional "black box" models. The proposed method achieves an accuracy of 87.5 % and 93.1 % for Uveal Melanoma and Cervical Cancer subtyping, respectively, which compares favorably to all competing approaches, including deep "black box" models. Summarizing a user study with pathologists, participants expressed positive feedback about the proposed explainable cancer subtyping pipeline due to its clinically-acceptable high accuracy and support for understandable and reliable explanations. Most importantly, the system's adherence to standard diagnostic procedures was perceived as a strength as it may facilitate integration into clinical practice.